"Edge-discovery batch 3 — event-driven (new data, higher bar)"

Date: 2026-09-10 | panel/split/costs as before. Events: 11,812 earnings (501 syms, dense 2020+) with tz-aware announcement stamps + surprise; FINRA short interest PIT via availableDate (2021-06+, 498 syms). DSR N=11 cumulative.

Pre-announcement window diagnostics: hit-rate(±15d)=0.898 (n=11238), drift→surprise spearman=0.026 (n=11029).

#Scoreboard

#H9 crowded-pead-short-squeeze -> RESEARCH

• OOS Sharpe 0.51 (IS 0.00), p vs 0 0.05, p vs PEAD0 1.00, corrPEAD0 0.90, beta 0.31
• v2: DSR 0.18 (N=11), PIT-200 0.46, lag2 0.51, tiered 0.49, pos-years 4/11
• cost ladder: optimistic=0.53, base=0.51, conservative=0.47, severe=0.40
• decay: H5=0.51; H10=0.54; H21=0.59; H42=0.59; H63=0.51
• yearly: 2016:0.00; 2017:0.00; 2018:0.00; 2019:0.00; 2020:0.00; 2021:-0.35; 2022:-0.46; 2023:1.22; 2024:1.53; 2025:0.95; 2026:1.27
• score parts: {'oos_sharpe': 6.4, 'signif': 14.3, 'regimes': 7.5, 'costs': 7.8, 'perturb': 10.0, 'breadth': 5.0, 'capacity': 5.0, 'decorr': 0.5, 'plaus': 3.0}

#H10 preannouncement-drift -> CANDIDATE

• OOS Sharpe 1.00 (IS 0.25), p vs 0 0.00, corrPEAD0 0.72, beta 0.33
• v2: DSR 0.51 (N=11), PIT-200 0.95, lag2 0.90, tiered 0.90, pos-years 8/11
• cost ladder: optimistic=1.08, base=1.00, conservative=0.76, severe=0.35
• decay: H5=0.91; H10=1.00; H21=0.58; H42=0.46; H63=0.77
• yearly: 2016:-0.85; 2017:1.22; 2018:-0.15; 2019:-0.00; 2020:0.71; 2021:1.54; 2022:0.35; 2023:1.52; 2024:2.21; 2025:1.12; 2026:2.85
• score parts: {'oos_sharpe': 12.5, 'signif': 15.0, 'regimes': 15.0, 'costs': 3.6, 'perturb': 10.0, 'breadth': 5.0, 'capacity': 4.9, 'decorr': 1.4, 'plaus': 2.5}

#H11 earnings-gap-fade -> RESEARCH

• OOS Sharpe -0.21 (IS 0.00), p vs 0 0.74, corrPEAD0 -0.01, beta -0.00
• v2: DSR 0.00 (N=11), PIT-200 0.00, lag2 -0.35, tiered -0.46, pos-years 4/11
• cost ladder: optimistic=-0.04, base=-0.21, conservative=-0.69, severe=-1.29
• decay: H5=-0.21; H10=-0.22; H21=-0.25; H42=0.00; H63=0.00
• yearly: 2016:0.00; 2017:0.00; 2018:0.00; 2019:0.00; 2020:-1.56; 2021:1.15; 2022:0.40; 2023:-3.00; 2024:-0.75; 2025:0.88; 2026:0.37
• score parts: {'oos_sharpe': 0.0, 'signif': 3.9, 'regimes': 0.0, 'costs': 0.0, 'perturb': 10.0, 'breadth': 5.0, 'capacity': 5.0, 'decorr': 4.9, 'plaus': 2.5}

#PEAD0 baseline (no promotion)

• OOS Sharpe 0.76; exists only to measure H9's short-interest increment.

#Reviewer addendum — hostile read

1. H9 is a falsified interaction, not a near-miss. The DTC gate turns 0.76 into 0.51 with p_vs_PEAD0 = 1.00 (strictly harmful), DSR 0.18, effective evidence only 2021+. Its grade-B score is component inflation (perturbation/cost subscores on a flat curve) — the decision-relevant statistics all say no. RESEARCH here means rejected.
2. H10 is promoted by rule but unproven by mechanism. The leakage story it was registered under is contradicted by its own diagnostic (drift→surprise spearman 0.026, n=11k). What survives is a pattern in search of an explanation — the leading alternative is announcement-premium capture (being long into scheduled news), which was NOT the tested hypothesis and needs its own test (e.g. does the return accrue pre- vs post-announcement? does it survive excluding high-surprise events?). Additional discounts: IS 0.25 (thin pre-2020 events), DSR 0.51 (coin-flip after 11 trials), corr 0.72 vs PEAD0 (shared exposure). Added to the daily paper log; any further step needs forward proof first.
3. H11 corroborates H4. Fading fails unconditionally (H4) AND event-conditionally (H11, with 2023 at -3.00). Stop fading gaps.
4. Sample honesty. All batch-3 evidence lives in 2020+ (earnings) and 2021+ (SI). Pre-2020 year rows are structural zeros, not weak years.
5. Cumulative funnel: 14 hypotheses tested (11 candidates + 3 baselines). 4 sit at CANDIDATE (H2/H5/H6/H10), all gated on forward proof. 0 have cleared untouched-holdout, paper, or shadow. Nothing is promotable today.

Runtime: 2.2 min.